Toward the Design of Chemical Libraries for Mass Screening Biased

Toward the Design of Chemical Libraries for Mass Screening Biased against Mutagenic Compounds ... Those factors are being brought into bear earlier in...
0 downloads 6 Views 179KB Size
J. Med. Chem. 2001, 44, 2793-2804

2793

Toward the Design of Chemical Libraries for Mass Screening Biased against Mutagenic Compounds Oriol Llorens† and Juan J. Perez Department of Chemical Engineering (UPC), The Barcelona School of Engineering, Av. Diagonal, 647, 08028 Barcelona, Spain

Hugo O. Villar* Telik, Inc., Chemoinformatics Group, 750 Gateway Boulevard, South San Francisco, California 94080 Received October 30, 2000

The ability to develop a chemical into a drug depends on multiple factors. Beyond potency and selectivity, ADME/PK and the toxicological profile of the compound play a significant role in its evaluation as a candidate for development. Those factors are being brought into bear earlier in the discovery process and even into the design of libraries for screening. The purpose of our study is the comparative analysis of simple physical characteristics of compounds that have been reported to be mutagens and nonmutagenic ones. The analysis of differences can lead to the development of knowledge-based biases in the libraries designed for massive screening. For each of four Salmonella strains, TA-98, TA-100, TA-1535, and TA-1537, an analysis of the statistical significance of the deviance of the averages for a number of global properties was carried out. The properties studied included parameters, such as topological indices, and bit strings representing the presence or absence of certain chemical moieties. The results suggest that mutagens display a larger number of hydrogen bond acceptor centers for most strains. Moreover, the use of bit strings points to the importance of certain molecular fragments, such a nitro groups, for the outcome of a mutagenicity study. Development of multivariate models based on global molecular properties or bit strings point to a small advantage of the latter for the prediction of mutagenicity. The benefits of the bit strings are in accord with the use of fragment-based approaches for the prediction of carcinogenicity and mutagenicity in methods described in the literature. Introduction The use of high-throughput screening has represented a major shift in the lead discovery process. Whenever the paradigm is applicable, the use of robotics and large chemical libraries has resulted in the identification of multiple compounds that could potentially serve as leads for medicinal chemistry programs for many targets. From a chemical perspective, the selection of compounds used for screening has gone through several stages of maturation. Initially, compounds were taken from historical collections or peptide libraries or were built around a single chemical scaffold, all of which produced little structural variety.1 The lack of variety spurred the development of techniques to assess the diversity of the libraries2 used, mainly with the purpose of removing the significant structural biases present. Multiple approaches to characterize the diversity of chemical libraries were put forward in the literature.3 Most of them use structural or physicochemical descriptors to evaluate similarity among the compounds, although other means have also been described.4 The massive screening of diversified libraries has yielded uneven results that are quite dependent on the families of targets but could result in the identification of multiple hits for an assay.5,6 In many cases, the hits * To whom correspondence should be addressed. Permanent address: Triad Therapeutics, 5820 Nancy Ridge Rd., Suite 200, San Diego, CA 92121. E-mail: [email protected]. Fax: (858) 455-0140. † Work carried out while visiting Telik, Inc.

found had to be subsequently discarded during the optimization stage because of their toxicological profile or poor bioavailability.6,7 The result has been a perception of misplaced effort that led to the latest stage in library design, with the tailoring of the libraries to suit their intended use as pharmacological agents.8 If compounds are ultimately to be used as drugs, certain biases are necessary, or at least helpful, to limit the range of properties of the compounds used for building up a library to those relevant to pharmaceuticals. Even within these constraints, maximal variability is still sought.9 The definition of the limits to be imposed on libraries has been part of active research.10-12 For the most part, the limits have been based on the physicochemical properties of known drugs. Significant attention has been paid to setting rules for the prediction of bioavailability or the ability to formulate the compounds into drug products.14-16 Toxicology is equally important in defining the viability of a chemical as a candidate for drug development. However, it is multifactorial and can only be approached in stages. One aspect of the toxicological package of a compound is its mutagenicity. A positive result in a mutagenicity test causes, as a rule, the discontinuation of work on that family of compounds. Consequently, the exclusion during the library design process of likely mutagens could help avoid wasteful screening of compounds and is consistent

10.1021/jm0004594 CCC: $20.00 © 2001 American Chemical Society Published on Web 07/20/2001

2794

Journal of Medicinal Chemistry, 2001, Vol. 44, No. 17

Llorens et al.

with the current trend to bring ADME/PK and toxicological considerations early on in the discovery process.1 For several decades, the Ames salmonella mutagenicity assay has been widely applied in research laboratories and by regulatory agencies as a useful tool for the detection of chemical mutagens found in the environment, natural products, food, pesticides, and drug candidates.17 The test represents one of the most widely used in vitro short-term screens for mutagenicity. Since the earlier application of the Ames test, different Salmonella strains have been generated to improve assay sensitivity and its range of applicability. These modifications have given the test enough versatility for the detection of different types of mutagenic compounds or compounds that become mutagenic upon metabolization by humans.18 For the development and maintenance of a reference data set, the use of standard tester strains has been recommended,19 although neither a gold standard nor a battery of strains are considered as fully satisfactory. Nevertheless, mutagenicity data in strains TA-98, TA100, TA-1535, and TA-1537 have been proposed in the literature as a starting screening routine.19 The classification of compounds as mutagenic or nonmutagenic results from combining the results of screening the different strains. Compounds are considered mutagenic if they give a positive result in any of the strains, with or without metabolic activation. In our analysis, the results for each strain will be independently scrutinized to capture different aspects of the mutagenic potential of a compound. The properties that each of the strains encapsulates are likely to be somewhat different, and some trends could be missed if all data sets were considered simultaneously. Unfortunately, the molecular mechanisms of mutagenic processes are not well characterized and are still under study. The lack of a molecular understanding complicates the development of structure-property relationships. Nonetheless, a significant amount of literature has been devoted over the years to the prediction of the mutagenicity and carcinogenicty by computational means.20 While the specifics of each approach vary greatly, methods for their prediction fall within two main camps: pattern recognition techniques or knowledge-based systems. The better known among the pattern recognition programs, CASE21 and its successor MULTICASE,22 TOPKAT,23 and ADAPT,24 predict toxicity based on the chemical structures. MULTICASE compares the distribution of molecular fragments found in a given molecule to those found in a database of fragments found in toxic substances. On the basis of the statistical weight of the compounds, it predicts toxicity or activity. Knowledge-based systems, such DEREK25 and HAZARDEXPERT,26 are based on the ability of the programs to identify molecular fragments or substructures that were previously known to give rise to various forms of toxicity. These programs constitute best efforts to accurately predict the activity of specific compounds and are extremely valuable later in the computer-assisted ligand design and optimization process. The purpose of this article is not the accurate prediction of the mutagenic potential of individual compounds, but the analysis of the properties and structural fea-

tures that can be found in mutagenic compounds and less frequently in nonmutagenic ones. The present work should be regarded as a first step toward the definition of molecular properties ranges, which could serve to introduce biases that enrich libraries in compounds with diminished genotoxic potential. Consequently, the types of properties that will be considered are the simple easily computed parameters that have been in use for library design, such as topological descriptors and bit strings. While global molecular properties provide a comprehensive view of a molecule, each bit in a bit string is associated to a certain molecular fragment. We will present the results of applying different statistical tools to investigate the biases present in libraries of compounds found to be active in the Ames test for each of the different Salmonella strains, with and without activation. Subsequently, we study the predictive power of each class of properties, as a means to evaluate the strategies that can be used during library design. Mutagenecity Data The CCRIS database28 was mined to select groups of active and inactive compounds in the Ames test for each of the four test strains recommended for use: TA-98, TA-100, TA-1535, and TA-1537. Results were retrieved for both: nonmetabolically activated compounds and compounds that underwent metabolic activation using rat liver S9 mix test. The data assembled was restricted to these two types to improve consistency. The classification of the results into positive (active) or negative (inactive) for each compound is not always straightforward because of ambiguities in the data reported by different laboratories. To minimize the uncertainties, only molecules that had a consistent rating in over 80% of the studies reported in the database were taken into consideration. CCRIS lacks structural information on the compounds; therefore, structures for the compounds were obtained by merging the data with the structures in the CMC and ACD (MDL Inc., San Leandro, CA) databases, using the CAS number as a linker which resulted in the database used for this study. This step further reduced the number of compounds available for study as not all entries could be successfully assigned a structure. For each molecule, two classes of molecular descriptors were utilized. On one side, topological indices together with other global molecular properties were computed; while on the other, a set of bit strings that has been described in the literature as particularly useful for diversity and similarity classifications,29 MDL’s MOLSKEYS, was used. First, a set of descriptors generated with the Molconn-X v 2.0 (Hall Associates, Quincy, MA) was computed for all the compounds in the database. These descriptors basically encode aspects of molecular size, shape, branching, and, to some extent, polarity. The property set monitored included formula weight (FW), number of rings (NRINGS), graph diameter (SDIAM), number of hydrogen bond donors (HBD) and acceptors (HBA), flexibility index (PHIA), Wiener number (W), Wiener p number (WP), total Wiener number (WT), Platt f number (PF), and Bonchev-Trinajstic indexes

Library Design for Screening against Mutagens

Journal of Medicinal Chemistry, 2001, Vol. 44, No. 17 2795

A Z test statistic is calculated to establish whether the two distributions are significantly different, as follows:

Table 1. Characteristics of Each of the Eight Data Sets Used for the Studya strain TA-98

metabolic activation activated not activated

TA-100

activated not activated

TA-1535 activated not activated TA-1537 activated not activated

activity

no. of compds

MW

C

N O

positive negative positive negative positive negative positive negative positive negative positive negative positive negative positive negative

335 1306 231 1708 322 1170 232 1557 93 838 61 1082 82 794 59 908

212.37 207.10 218.92 209.13 201.16 209.70 199.36 212.45 173.33 206.95 183.23 208.59 226.19 210.84 231.59 213.06

11 10 10 10 10 10 8 11 8 10 7 10 13 10 12 10

2 1 2 1 1 1 2 1 1 1 1 1 1 1 2 1

2 2 3 2 2 2 3 2 2 2 2 2 2 2 3 2

a MW is the average molecular weight, while C, N, and O show the average count of those atoms in the data set.

(SIDC, SIDW). The octanol-water partition coefficient (LogP) was also computed based on the ACD software (ACD, Ontario Canada). The second set of descriptors was MDL’s MOLSKEYS.32 Each of the 166 bits encodes the presence or absence of determined substructural patterns. Those MOLSKEYS are automatically created by the ISIS package (MDL, Inc., San Leandro, CA) at the time of the generation of a database. Each substructure bit was treated during the statistical analysis as a separate binary descriptor. All statistical calculations were carried out using the algorithms embedded in the S-Plus 200 package (MathSci, Bothell, WA). Recursive partitioning and Wilcoxon rank-sum test were carried out as embodied in the S-PLUS “tree” and “wilcox.test” functions, respectively. Graphics were generated using S-PLUS capabilities. Statistical Analysis of the Properties Table 1 shows the general characteristics of each of the databases generated for this study and shows that they are comparable in their gross characteristics. The data sets are divided by Salmonella strain and whether the compounds have been metabolically activated, generating a total of eight groups to be analyzed independently. The strains TA-98 and TA-100 display a similar frequency of positive compounds, while the number is much smaller for the TA-1535 and TA-1537. In every other respect, the gross molecular characteristics of the compounds that compose each data set appear comparable. The Wilcoxon rank sum test for two independent variables was carried out for each of the global properties computed to determine if statistically significant differences existed in the properties of positive and negative compounds. The Wilcoxon rank-sum test is widely used to analyze the differences between independent and not-paired data sets, taking into account their magnitude.30 The test ranks all data assigning a value of 1 for the lowest value increasingly to the largest one, then detects how these rankings are distributed between the two groups in a way that if high or low ranks are found predominantly in one sample, this means the two populations are not identical.

Z ) (R - µR)/σR where

µR ) ((n1n2(n1 + n2 + 1))/2)1/2 n1 is the size of the reference sample; n2 is the size of the sample for comparison; R is sum of ranks of the reference sample; σR is the standard deviation of R. At 95% significance, the critical Z values for a twotailed test are (1.96.30 Values of Z outside the critical region lead to the rejection of the null hypothesis, and the two distributions are considered different. The Wilcoxon test has several advantages for data analysis over other conventional statistical tests. The most convenient aspects of the test are that the data does not need to follow a normal distribution, and the sizes of the populations do not need to be equal. (a) Global Molecular Properties. Figure 1 shows a box plot of the global properties used for positive and negative compounds for each of the eight data sets. The box plot shows the median as a stripe, the upper and lower quartiles of the data distribution as the box, and the whiskers show the extent of the data beyond the quartiles. The box plot allows a rapid visualization of the data sets. For example, if the upper and lower quartiles of the box plot are at about the same distance from the median stripe, the data is distributed symmetrically around the stripe in the box. The results of the Wilcoxon analysis for the global properties are also part of Figure 1. The Z test data can be analyzed in three different ways taking into account (i) the statistical significance of the differences observed; (ii) the sign of Z, and (iii) the magnitude of Z. The two data sets are considered to be different if |Z| > 1.96. Positive Z values indicate that the negative compounds show higher values for that descriptor. Larger absolute values for Z point to more significant differences between the two distributions. The data shows that clear, statistically significant differences exist in the properties of positive and negative compounds for each of the strains. It is also interesting that the properties of the compounds found to be positive in each strain are somewhat distinct. The ability to form hydrogen bonds is, in the case of the molecules that did not undergo activation, a clear discriminating factor. Compounds found to be positive with the TA-98, TA-100, and TA-1537 strains have a larger number of hydrogen bond acceptors, while the number of hydrogen bond donor centers is an important discriminant for TA-98 and TA-100. The hydrophobicity of the compounds as characterized by their calculated LogP, has no bearing on the activity of the compounds for the TA-98 strain. This is a somewhat surprising result, because of the emphasis that has been given to the cLogP in the prediction of mutagenicity.31 Nevertheless, the hydrophobic character of the molecule has influence for TA-100 and TA-1537. For the TA-100, the positive compounds are less lipophilic, while the opposite is true for the TA-1537 strain.

2796

Journal of Medicinal Chemistry, 2001, Vol. 44, No. 17

Llorens et al.

Library Design for Screening against Mutagens

Journal of Medicinal Chemistry, 2001, Vol. 44, No. 17 2797

Figure 1. Box plots for each of the properties and the strains studied: (a) compounds did not undergo activation; (b) with the compounds undergoing metabolic activation. P stands for compounds giving a positive response in the assay, while N is for the negatives. The stripe in the box indicates the value of the mean.

2798

Journal of Medicinal Chemistry, 2001, Vol. 44, No. 17

Llorens et al.

Library Design for Screening against Mutagens

Journal of Medicinal Chemistry, 2001, Vol. 44, No. 17 2799

Figure 2. Frequency of the distribution of each ISIS/MOLSKEYS for each strain with or without activation.

2800

Journal of Medicinal Chemistry, 2001, Vol. 44, No. 17

Table 2. Identity of Certain Particular Bits in ISIS-MOLSKEYSa bit no.

key

bit no.

key

16 24 57 70 84 105 114 116 122 127 130 133 143 149 153 156 160

QAA@1 N-O O in heterocycle QNQ NH2 A$A($A)$A CH3CH2A CH3AACH2A AN(A)A A$A! > 1 QQ > 1 A$A!N A$A!O CH3 > 1 QCH2A NA(A)A CH3

22 52 63 71 87 109 115 118 125 128 132 141 144 150 155 157 162

three-membered ring NN NdO NO X!A$A A CH2O CH3ACH2A ACH2CH2A > 1 aromatic rings > 1 ACH2AAACH2A OACH2A CH3 > 2 and other rare features A-not%-A%A-not%-A A!A$A!A A!CH2!A C-O aromatic

a Q stands for any atom other than C or H, A any atom other than H. $ represents a bond in a Ring, ! a bond in a chain, @n atom n attached to cycle.

These results may indicate that there is a window of hydrophobiticity that nonmutagenic compounds fall within. When the compounds are activated using the rat liver S9 mix assay, the properties that distinguish between positive and negative compounds are not the same. The number of rings is a factor that shows a different distribution in positive activated compounds, the TA98 and TA-100 strains being larger among the positive compounds. The number of hydrogen bond donors is quite significant as a discriminant for all strains but TA-1535, which appears to be dependent on more subtle factors that are better captured by topological parameters such as the Wiener index. A comparison of the parameters that show divergence between positive and negative compounds makes it clear that the compounds that are detected as mutagenic by each strain have different physicochemical properties. The most significant is that the average number of hydrogen bond donors or acceptor groups appears to be larger for the positive compounds. (b) Bit String Properties and Fragment Analysis. The bar plots in Figure 2 show the difference in the frequency with which each of the 166 bits described by the MOLSKEYS appears between positive and negative compounds. The sign of that bar plot determines whether the particular bit string is more abundant in the positive or in the negative set of compounds, respectively. Table 2 shows some of the MOLSKEYS of interest to this study and their associated moieties. While some bits appear with similar frequency in positive and negative compounds, others show a marked preference for either kind of compounds. The result in itself could be expected, because bit strings are associated to the presence of certain substructures or moieties in chemicals, and fragment-based methods are the basis of techniques described to predict the mutagenic character of compounds. The bar plots show that for the TA98 and TA-100 strains keys 24, 63, 70, 71, and 122, related with NO2 and NO groups, are overrepresented. All these keys are strongly paired with positive activity. Interestingly, some bits appear less frequently in the positive compounds. The implication would be that certain fragments might actually decrease the potential

Llorens et al.

that a given compound would be positive in the Ames test. Keys 149, 157, and 160, encoding substructures related with aliphatic character and carboxyl groups, respectively, are most common in negative compounds. Strain TA-1537 shows an overrepresentation of bits numbered 105, 125, and 127 and 141, encoding aromatic rings and heteroatoms in a cycle strongly related with activity. Conversely, for the same strain, the presence of alkyl chains (MOLSKEYS 153 and 155) is again associated with the negative compounds. Finally, TA1535 presents MOLSKEYS 70 and 130, encoding the presence of heteroatoms related with positive activity. Surprisingly, aromaticity (MOLSKEY 162) is correlated with negative activity for this strain. As it was the case for global properties, each strain is responding to compounds containing different molecular fragments. TA-98 and TA-100 have the closest similarity in the fragments that are overrepresented, but still the two strains are sufficiently different to be able to discriminate among them. On the basis of the results obtained from the bits overrepresented in the MOLSKEYS, a subset of chemical functionalities was selected to determine whether they were at least in part responsible for the higher frequency of particular bit strings. Table 3 shows the results of searching for the presence of specific substructural moieties in positive compounds in assays carried out with and without activation for each of the bit strings. The table shows the number of positive and negative molecules containing each fragment for each strain and assay, as well as the binomial probability (p binomial) that the distribution could be due to chance. The lower the binomial probabilities, the more significant the difference in the frequency of appearance of that functional group or moiety compared to what could be expected if it was a random event. For TA-1535 and TA-1537, the limited number of positive compounds does not permit achieving statistical significance in some cases where a bias appears to be present. Again, different strains discriminate the mutagenic potential of distinct chemical classes. TA-98 and TA100 show some degree of correlation because, for both strains, PhNO2 groups and N ) X groups are disproportionately represented among the positive. TA-1535 identified epoxy and haloalkanes compounds as mutagenic, but interestingly all other strains show only a slight bias against these compounds. Certain functional groups appear to be less likely to be positive against any strain in assays with or without activation. For instance, compounds containing carboxylic, propyl, and even methyl groups are less frequent in positive compounds for any of the strains, compared to what could be expected considering their abundance in the overall library. Results in Table 3 are consistent with previous structural alerts about mutagenicity or carcinogenicity.32 While there are global properties and fragments that can be found as biased in all assays, in a majority of cases, the results from each strain are different. Mutagenicity Prediction While our aim is not the prediction of the activity of individual mutagenic compounds, whether the properties discussed are indeed able to discriminate active from inactive compounds is important in their evalua-

Library Design for Screening against Mutagens

Journal of Medicinal Chemistry, 2001, Vol. 44, No. 17 2801

Table 3. Molecular Fragments Found in Positive (P) and Negative (N) Compounds, Based on the Results for Compounds without and with Metabolic Activation Without Metabolic Activation TA-98

TA-100

TA-1535

moieties

P (231)

N (1708)

p

P (232)

N (1789)

p

P (61)

N (1082)

PhNH2 PhNO2 Ph-O-C Ph-S-C C-CH2X C-CH2-O-CH2-C HO-N-C2 Ph Ph C-COOH NdC, NdO, NdN CH2-CH2-CH2 C-CH3 epoxy sulfonamide methylsulfoxide -NH2 -NHC C-CO-NCC

34 118 20 5 8 1 0 69 21 149 5 47 5 0 0 6 0 4

136 106 176 21 76 36 1 317 339 263 244 788 22 21 11 23 5 33

0.05 >0.05

17 89 13 4 28 1 0 42 27 123 8 46 15 1 1 5 4 3

161 112 165 24 37 30 1 328 285 258 229 722 15 22 9 28 21 26